Pseudorandom
Context
A Rhino plugin for architects to render with AI using material prompts
Pseudorandom lets users describe what they want in plain language, attach prompts to objects, and generate images with respect to spatial context. Additionally, it allows teams to use their own custom rendering workflows.
Problem
Architects using third-party AI workflows face legal and ethical risks
Baked into community-made ComfyUI workflows are models potentially trained on undisclosed, copyrighted, or nonconsensual datasets, exposing firms to client breach of contract and legal liability.
Earlier iterations of provenance tracking relied on a single free-text field where workflow creators could type whatever they wanted about a model's origin.
Solution
The Pseudorandom Vetting Framework
Instead of asking architects to guess if a model is safe, we built an open, centralized provenance repository hosted directly on Hugging Face.
Pseudorandom researches each model and maintains uneditable records that quantify safety across three core metrics:
Live model repositories on Pseudotools Hugging Face
This repository serves as the system's single source of truth, dynamically feeding provenance data directly into ComfyUI, Rhino, and final render outputs.
User Flow
An End-to-End Provenance Pipeline
I designed for the provenance details to be carried through the Hugging Face repositories, creating your own workflow in ComfyUI, and even after rendering the image in Rhino.
(Pre-render)
Model Database
Architects browse Pseudorandom-vetted models, review provenance records, and download verified assets directly.
Pseudocomfy Custom Nodes
Within ComfyUI, users author workflows using custom model loader nodes that pull models via direct reference to their Hugging Face records.
Workflow Converter
A utility tool scans raw ComfyUI workflows, detects Pseudocomfy nodes, and packages the required metadata payload automatically.
(At-render)
Workflow Information Popouts
Inside Rhino, selecting a workflow displays an aggregated vetting badge. Clicking the popout expands detailed metadata on individual models, licenses, and author attribution before rendering begins.
(Post-render)
Image Lineage
Provenance metadata is embedded directly into rendered image files. Architects can drop outputs into the Image Inspector to verify model lineage for client handoffs or compliance records.
Impact
Shipped, not theoretical
The model database is live, built on these new records. The packaging step and the image viewer are both shipped and working. A public exhibition is planned for February.
Reflection
What I've learned
Emotion is subjective
Mapping physical inputs to AI audio parameters revealed that dynamic moods require dynamic logic. With more time, we would refine how the system prompt interprets edge-case slider combinations to ensure every output feels cohesive.
Designing for feeling
Evaluating our work on how it makes people feel kept us grounded. Holding onto that empathy allowed us to stay curious and open as the physical form factor evolved from desk to wall.